AI Coach
Finding the real need behind "pronunciation practice." A survey said users wanted pronunciation practice. I doubted the literal reading, and follow-up research confirmed it: most meant fluency, word choice and confidence. That reframe became the AI Coach.
- ~80%meant fluency, not phonetics
- 92%rated the prototype useful
- 77%felt more confident after a session
- 7interviews, plus a 26-person survey
The project
- Role
- Product Designer + Product Manager
- Timeline
- Jun '25 (survey) → Aug '25 (interviews) → Oct '25 (PRD) → Nov/Dec '25 (beta launch)
- Company
- RealLife English, B2C ESL app
- Team
- CEO (AI sponsor), Product Lead (pricing and packaging), Lead Engineer (AI infrastructure)
- Methods
- Importance/satisfaction survey, moderated interviews, live prototype, RevenueCat
Executive summary
An importance-vs-satisfaction survey said users wanted "pronunciation practice." I doubted the literal reading, followed up with qualitative and quantitative research, and confirmed it: roughly 80% of users meant broader speaking fluency, word choice and confidence, not phonetics. That reframe became the AI Coach concept: watch a short lesson, speak about the vocabulary you just learned in context, then get feedback.
What this case does not claim: a measured business outcome. A churn number that looked like a win didn't survive segmentation, because the feature launched together with a Black Friday campaign. Finding that out is part of the case, not hidden from it.
This is not an inflated case. It's a "the right thing done the right way" story, with external influences outside my control.
PM signals
- Caught a false signal in quant data before it became a mis-scoped feature. The doubt was recorded when the survey came in, then confirmed by dedicated follow-up research. Not a story reconstructed after the fact.
- Quant → quali triangulation as a working method. A 26-person survey and 7 interviews were designed to re-test the meaning of one ambiguous word before it shaped a feature spec.
- Refused to claim an effect the data doesn't support. Walked back a churn figure after segmenting monthly vs. yearly plans: the "improvement" was a plan-mix artifact, not a product effect.
- Operated inside mandate boundaries. Pricing and packaging decisions (a one-call lifetime free cap, Black Friday bundling) sat outside my scope, and are the most likely reason the feature's real impact was never measurable.
Hands-on: product designer hybrid
After the research, I designed the UI and UX of the new feature.
The same gap, solved solo: AI product demoFull technical freedom, no organizational constraints: prompts, evals and adversarial tests.